Professional Documents
Culture Documents
S1566-2535(13)00047-X
http://dx.doi.org/10.1016/j.inffus.2013.04.006
INFFUS 563
To appear in:
Information Fusion
Please cite this article as: M. Woniak, M. Graa, E. Corchado, A Survey of Multiple Classifier Systems as Hybrid
Systems, Information Fusion (2013), doi: http://dx.doi.org/10.1016/j.inffus.2013.04.006
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers
we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and
review of the resulting proof before it is published in its final form. Please note that during the production process
errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
b Computational
Abstract
A current focus of intense research in pattern classification is the combination of several classifier systems, which can be built following either the same or different models and/or datasets
building approaches. These systems perform information fusion of classification decisions at
different levels overcoming limitations of traditional approaches based on single classifiers. This
paper presents an up-to-date survey on multiple classifier system (MCS) from the point of view
of Hybrid Intelligent Systems. The article discusses major issues, such as diversity and decision fusion methods, providing a vision of the spectrum of applications that are currently being
developed.
Keywords: hybrid systems, hybrid classifier, multiple classifier system, machine learning,
combined classifier, pattern classification.
1. Introduction
Hybrid Intelligent Systems offer many alternatives for unorthodox handling of realistic increasingly complex problems, involving ambiguity, uncertainty and high-dimensionality of data.
They allow to use both a priori knowledge and raw data to compose innovative solutions. Therefore, there is growing attention to this multidisciplinary research field in the computer engineering research community. Hybridization appears in many domains of human activity. It has an immediate natural inspiration in the human biological systems, such as the Central Nervous System,
which is a de facto hybrid composition of many diverse computational units, as discussed since
the early days of computer science e.g., by von Neumann [1] or Newell [2]. Hybrid approaches
seek to exploit the strengths of the individual components, obtaining enhanced performance by
their combination. The famous no free lunch theorem [3] stated by Wolpert may be extrapolated to the point of saying that there is no single computational view that solves all problems.
Fig. 1 is a rough representation of the computational domains covered by the Hybrid Intelligent
System approach. Some of them deal with the uncertainty and ambiguity in the data by probabilistic or fuzzy representations and feature extraction. Others deal with optimization problems
appearing in many facets of the intelligent system design and problem solving, either following a
Corresponding
author
Email addresses: michal.wozniak@pwr.wroc.pl (Micha Wozniak), ccpgrrom@gmail.com (Manuel Grana),
escorchado@usal.es (Emilio Corchado)
Preprint submitted to Information Fusion
nature inspired or a stochastic process approach. Finally, classifiers implementing the intelligent
decision process are also subject to hybridization by various forms of combination. In this paper,
we focus in this specific domain, which is in an extraordinary effervescence nowadays, under
the heading of Multi-Classifier Systems (MCS). Referring to classification problems, Wolperts
theorem has an specific lecture: there is not a single classifier modeling approach which is optimal for all pattern recognition tasks, since each has its own domain of competence. For a given
classification task, we expect the MCS to exploit the strengths of the individual classifier models at our disposal to produce the high quality compound recognition system overcoming the
performance of individual classifiers. Summarizing:
Hybrid Intelligent Systems (HIS) are free combinations of computational intelligence techniques to solve a given problem, covering al computational phases from data normalization up to final decision making. Specifically, they mix heterogeneous fundamental views
blending them into one effective working system.
Information Fusion covers the ways to combine information sources in a view providing
new properties that may allow to solve better or more efficiently the proposed problem.
Information sources can be the result of additional computational processes.
Multi-Classifier Systems (MCS) focus on the combination of classifiers form heterogenous
or homogeneous modeling backgrounds to give the final decision. MCS are therefore a
subcategory of HIS.
NATUREINSPIRED
SYSTEMS
HYBRID OPTIMIZATION
METHODS
COMBINED
CLASSIFIERS
UNCERTAINTY
MANAGEMENT
Historical perspective. The concept of MCS was first presented by Chow [4], who gave conditions for optimality of the joint decision1 of independent binary classifiers with appropriately
defined weights. In 1979 Dasarathy and Sheela combined a linear classifier and one k-NN classifier [6], suggesting to identify the region of the feature space where the classifiers disagree.
The k-NN classifier gives the answer of the MCS for the objects coming from the conflictive
1 We can retrace decision combination long way back in history. Perhaps the first worthy reference is the Greek
democracy (meaning government of the people) ruling that full citizens have an equal say in any decision that affects
their life. Greeks believed in the community wisdom, meaning that the rule of the majority will produce the optimal
joint decision. In 1785 Condorcet formulated the Jury Theorem about the misclassification probability of a group of
independent voters [5]], providing the first result measuring the quality of classifier committee.
Figure 2: Evolution of the number of publications per year ranges retrieved from the keywords specified in the plot
legend. Each plot corresponds to searching site: the top to Google Scholar; the center to the Web of Knowledge, the
bottom to Scopus. The first entry of the plots is for publications prior to 1990. The last entry is only for the last two
years.
region and by the linear one for the remaining objects. Such strategy significantly decreases the
exploitation cost of whole classifier system. This was the first work introducing a classifier selection concept, however the same idea was developed independently in 1981 by by Rastrigin and
Erenstein [7] performing first a feature space partitioning and, second, assigning to each partition
region an individual classifier that achieves the best classification accuracy over it. Other early
relevant works formulated conclusions regarding MCS s classification quality, such as [8] who
considered a neural network ensemble, [9] with majority voting applied to handwriting recognition, Turner in 1996 [10] showed that averaging outputs of an infinite number of unbiased
and independent classifiers can lead to the same response as the optimal Bayes classifier, Ho
[11] underlined that a decision combination function must receive useful representation of each
classifiers decision. Specifically, they considered several method based on decision ranks, such
as Borda count. Finally, the landmark works devoted introducing bagging [12] and boosting
[13; 14] which are able to produce strong classifiers [15], in the (Probably Approximately Correct) theory [16] sense, on the basis of the weak one. Nowadays MCS, are highlighted by review
articles as a hot topic and promising trend in pattern recognition [1721]. These reviews include
the books by Kuncheva [22], Rokach [23], Seni and Edler [24], and Baruque and Corchado [25].
Even leading-edge general machine learning handbooks such as [2628] include extensive presentations of MCS concepts and architectures. The popularity of this approach is confirmed by
the growing trend in the number of publications shown in Fig. 2. The figure reproduces the
evolution of the number of references retrieved by the application of specific keywords related
to MCS since 1990. The experiment was repeated on three well known academic search sites.
The growth in the number of publications has an exponential trend. The last entry of the plots
corresponds to the last two years, and some of the keywords give as many references as in the
previous five years.
Advantages. Dietterich [29] summarized the benefits of MCS: (a) allowing to filter out hypothesis that, though accurate, might be incorrect due to a small training set, (b) combining classifiers
trained starting from different initial conditions could overcome the local optima problem, and
(c) the true function may be impossible to be modeled by any single hypothesis, but combinations of hypotheses may expand the space of representable functions. Rephrasing it, there is
widespread acknowledgment of the following advantages of MCS:
MCS behave well in the two extreme cases of data availability: when we have very scarce
data samples for learning, and when we have a huge amount of them at our disposal. In
the scarcity case, MCS can exploit bootstrapping methods, such as bagging or boosting.
Intuitive reasoning justifies that the worst classifier would be out of the selection by this
method [30] e.g., by individual classifier output averaging [31]. In the event of availability
of a huge amount of learning data samples, MCS allow to train classifiers on datasets
partitions and merge their decision using appropriate combination rule [20].
Combined classifier can outperform the best individual classifier [32]. Under some conditions (e.g., majority voting by a group of independent classifiers) this improvement has
been proven analytically [10].
Many machine learning algorithms are de f acto heuristic search algorithms. For example the popular decision tree induction method C4.5 [33] uses a greedy search approach,
choosing the search direction according to an heuristic attribute evaluation function. Such
an approach does not assure an optimal solution. Thus, the combined algorithm, which
4
could start its work from different initial points of the search space, is equivalent to a multistart local random search which increases the probability of finding an optimal model.
MCS can easily be implemented in efficient computing environments such as parallel and
multithreaded computer architectures [34]. Another attractive area of implementation solutions is distributed computing systems (i.e.: P2P, Grid or Cloud computing) [35; 36],
especially when a database is partitioned for privacy reasons [37] so that partial solutions
must be computed on each partition and only the final decision is available as the combination of the networked decision.
Wolpert stated that each classifier has its specific competence domain [3] where they overcome other competing algorithms, thus it is not possible to design a single classifier which
outperforms another ones for each classification tasks. MCS try to select always the local
optimal model from the available pool of trained classifiers.
System structure. The general structure of MCS is depicted in Fig. 3 following a classical pattern
recognition [38] application structure. The most informative or discriminant features describing
the objects are input to the classifier ensemble, formed by a set of complementary and diverse
classifiers. An appropriate fusion method combines the individual classifier outputs optimally
to provide the system decision. According to Ho [39], two main MCS design approaches can
be distinguished. On one hand, the so-called coverage optimization approach tries to cover the
space of possible models by the generation of a set of mutually complementary classifiers whose
combination provides optimal accuracy. On the other hand, the so-called decision optimization
approach concentrates on designing and training an appropriate decision combination function
over a set of individual classifier given in advance [40].The main issues in MCS design are:
System topology: How to interconnect individual classifiers,
Ensemble design: How to drive the generation and selection of a pool of valuable classifiers,
Fuser design: How to build a decision combination function (fuser) which can exploit the
strengths of the selected classifiers and combine them optimally.
ensemble
classifer 1
object
classifier 2
fuser
classifier n
decision
2. System topology
Fig. 4 illustrates the two canonical topologies employed in MCS design. The overwhelming
majority of MCS reported in the literature is structured in a parallel topology [22]. In this architecture, each classifier is feed the same input data, so that the final decision of the combined
classifier output is made on the basis of the outputs of the individual classifiers obtained independently. Alternatively, in the serial (or conditional) topology, individual classifiers are applied
in sequence, implying some kind of ranking or ordering over them. When the primary classifier
can not be trusted to classify a given object e.g., because of the low support/confidence in its
result, then the data is feed to a secondary classifier [41; 42], and so on, adding classifiers in
sequence. This topology is adequate when the cost of classifier exploitation is important, so that
the primary classifier is the computationally cheapest one, and secondary classifiers have higher
exploitation cost [43]. This model can be applied to classifiers with the so-called reject option
as well [44]. In [45] the first classifier in the pipeline gives an estimation of the certainty of the
classification, so that uncertain data samples are sent to a second classifier, specialized in difficult
instances. We notice the similarity of such approach to the ordered set of rules [46] or decision
list [47], when we consider each rule as the classifier.
A very special case of sequential topology is the Adaboost introduced by Freund and Schapire
in 1995 [48], widely applied in data mining problems [49]. The goal of boosting is to enhance
the accuracy of any given learning algorithm, even weak learning algorithms with an accuracy
slightly better than chance. Shapire [50] showed that weak learners can be boosted into a strong
learning algorithm by sequentially focusing on the subset of the training data that is hardest
to classify. The algorithms performs training of the weak learner multiple times, each time
presenting it with an updated distribution over the training examples. The distribution is altered
so that hard parts of the feature space have higher probability, i.e. trying to achieve a hard
margin distribution. The decisions generated by the weak learners are combined into a final
single decision. The novelty of Adaboost lies in the adaptability of the successive distributions
to the results of the previous weak learners, thus the name AdaptiveBoost. In the words of
Kivinen et. al [51], AdaBoost finds a new distribution that is closest to the old one but taking into
consideration the restriction that the new distribution must be orthogonal to the mistake vector
of the current weak learner.
classifer 1
classifier 2
object
fuzer
decision
classifier n
object
object
classifer 1
classifer 1
classifer 2
classifer 2
classifer n
classifer n
decision
decision
3. Ensemble design
Viewing MCS as a case of robust software [5255], diversity arises as the guiding measure
of the design process. Classifier ensemble design aims to include mutually complementary individual classifiers which are characterized by high diversity and accuracy [56]. The emphasis
from the Hybrid Intelligent System point of view is in building MCS from components following different kinds of modeling and learning approaches, expecting an increase in diversity and
a decrease in classifier output correlation [57]. Unfortunately, the problem of how to measure
classifier diversity is still an open research topic. Brown et al. [58] notice that we can ensure
diversity using implicit or explicit approaches. Implicit approaches include techniques of independent generation of individual classifiers, often based on random techniques, while explicit
approaches focus on the optimization of a diversity metric over a given ensemble line-up. In this
second kind of approaches, individual classifier training is performed conditional to the previous
classifiers with the aim of exploiting the strengths of valuable members of classifier pool. This
section discusses some diversity measures, and the procedures followed to ensure diversity in the
ensemble.
3.1. Diversity measures
For regression problems, the variance of the outputs of ensemble members is a convenient
diversity measure, because it was proved that the error of a compound model based on a weighted
averaging of individual model outputs can be reduced according to increasing diversity [56;
59]. Brown et al. [60] showed a functional relation between diversity and individual regressor
accuracy, allowing to control the bias-variance tradeoff systematically.
For classification problems such theoretical results have not been proved yet, however many
diversity measures have been proposed till now. On the one hand, it is intuitive that increasing
diversity should lead to the better accuracy of the combined system, but there is no formal proof
of this dependency [61], as confirmed by the wide range of experimental results presented e.g., in
[62]. In [53] authors decomposed the error of the classification by majority voting into individual
accuracy, good and bad diversities. The good diversity has positive impact on ensemble error
reduction, whereas the bad diversity has the opposite effect. Sharkley et al. [55] proposed a
hierarchy of four levels of diversity according to the answer of the majority rule, coincident
failures, and possibility of at least one correct answer of ensemble members. Brown et al. [58]
argue that this hierarchy is not appropriate when the ensemble diversity varies between feature
subspaces. They formulated the following taxonomy of diversity measures:
Pairwise measures averaging a measure between each classifier pair in an ensemble, such
as Q-statistic [58], kappa-statistics [63], disagreement [64] and double-fault measure [61;
65].
Non-pairwise diversity measures comparing outputs of a given classifier and the entire ensemble, such as Kohavi-Wolpert variance [66], a measure of inter-rater (inter-classifier)
reliability [67], the entropy measure [68], the measure of difficulty [8], generalized diversity [52], and coincident failure diversity [69].
The analysis of several diversity measures [70] relating them to the concept of classifiers margin, showed their limitations and the source of confusing empirical results. They relate the classifier selection to a NP-complete matrix cover problem, implying that ensemble design in fact a
7
quite difficult combinatorial problem. Diversity measures usually employ the most valuable subensemble in ensemble pruning processes [71]. To deal with the high computational complexity
of ensemble pruning, several hybrid approaches have been proposed such as heuristic techniques
[72; 73], evolutionary algorithms [74; 75], reinforcement learning [76], and competitive crossvalidation techniques [77]. For classification tasks, the cost of acquiring feature values (which
could be interpreted as the price for examination or time required to collect the data for decision
making) can be critical. Some authors take it into consideration during the component classifier
selection step [78; 79].
3.2. Ensuring Diversity
According to [22; 38] we can enforce the diversity of a classifier pool by the manipulation of
either individual classifier inputs, outputs, or models.
3.2.1. Diversifying input data
This diversification strategy assumes that classifiers trained on different (disjoint) input subspaces become complementary. Three general strategies are identified:
1. Using different data partitions.
2. Using different sets of features.
3. Taking into consideration the local specialization of individual classifiers.
Data partitions. They may be compelled by several reasons, such as data privacy, or the need to
learn over distributed data chunks stored in different databases [8082]. Regarding data privacy,
we should notice that using distributed data may come up against legal or commercial constraints
which do not allow sharing raw datasets and merging them into a common repository [37]. To ensure privacy we can train individual classifiers on each database independently and merge their
outputs using hybrid classifier principles [83]. The distributed data paradigm is strongly connected with the big data analysis problem [84]. A huge database may impede to deliver trained
classifiers under specified time constraints, imposing to resort to sampling techniques to obtain
manageable dataset partitions. A well known approach is cross-validated committee which requires to minimize overlapping of dataset partitions [56]. Providing individualized train datasets
for each classifier is convenient in the case of shortage of learning examples. Most popular techniques, such as bagging [12] or boosting [14; 19; 64; 85], have their origin in bootstrapping [13].
These methods try to ascertain if a set of weak classifier may produce a strong one. Bagging
applies sampling with replacement to obtain independent training datasets for each individual
classifier. Boosting modifies the input data distribution perceived by each classifier from the results of classifiers trained before, focusing on difficult samples, making the final decision by a
weighted voting rule.
Data features. May be selected to ensure diversity training of a pool of classifiers. The Random
Subspace [86; 87] was employed for several types of the individual classifiers such as decision
tree (Random Forest) [88], linear classifiers [89], or minimal distance classifier [90; 91]. It is
worth pointing out the interesting propositions dedicated one-class classifier presented by Nanni
[92] or an hierarchical method of ensemble forming, based on feature space splitting and then
assigning two-class classifiers (i.e. Support Vector Machines) locally presented in [93; 94].
Attribute Bagging [95] is a wrapper method that establishes the appropriate size of a feature
subset, and then creates random projections of a given training set by random selection of feature
subsets. The classifier ensemble are train on the basis of the obtained set.
8
Local specialization. It is assumed for classifier selection, selecting the best single classifier
from a pool of classifiers trained over each partition of the feature space. It gives the MCS answer for all objects included in the partition [7]. Some proposals assume classifier local specialization, providing only locally optimal solutions [38; 72; 9698], while others divide the feature
space, selecting (or training) a classifier for each partition. Static and dynamic approaches are
distinguished:
Static classifier selection [99]: the relation between region of competence and assigned
classifier is fixed. Kunchevas Clustering and Selection algorithm[100] partitions the feature space by a clustering algorithm, and selects the best individual classifier for each
cluster according to its local accuracy. Adaptive Splitting and Selection algorithm in [101]
partitions the feature space and assigns classifiers to each partition into one integrated
process. The main advantage of AdaSS is that the training algorithm considers an area
contour to determine the classifier content and, conversely, that the region shapes adapt to
the competencies of the classifiers. Additionally, the majority voting or more sophisticated
rules are proposed as combination method of area classifiers [102]. Lee et al. [103] used
the fuzzy entropy measure to partition the feature space and select the relevant features
with good separability for each of them.
Dynamic classifier selection: the competencies of the individual classifiers are calculated
during classification operation [104107]. There are several interesting proposals which
extend this concept e.g., by using preselected committee of the individual classifier and
making the final decision on the basis of a voting rule [108]. In [109; 110] authors propose
dynamic ensemble selection based on the original competence measure using classification
of so-called random reference classifier.
Both static [111113] and dynamic [114116] classifier specialization are widely used for data
stream classification.
3.2.2. Diversifying outputs
MCS diversity can be enforced by the manipulation of the individual classifier outputs, so
that an individual classifier is designed to classify only some classes in the problem.
The combination method should restore the whole class label set e.g., a multi-class classification problem can be decomposed into a set of binary classification problems [117; 118]. The
most popular propositions of two-class classifier combinations are: OAO (one-against-one) and
OAA (one-against-all)[119], where at least one predictor relates to each class. The model that
a given object belongs to a chosen class is tested against the alternative of the feature vector
belonging to any other class. In the OAA method, a classifier is trained to separate a chosen
class from the remaining ones. OAA returns class with maximum support. In more general
approaches, the combination of individual outputs is made by finding the closest class, in some
sense, to the code given by the outputs of the individual classifiers. ECOC (Error Correcting
Output Codes) model was proposed by Dieterich and Bakiri [118], who assumed that a set of
classifiers produces sequence of bits which is related to code-words during training. The ECOC
points at the class with the smallest Hamming distance to its codeword. Passerini et al. showed
advantages of this method over traditional ones for the ensemble of support vector machines
[120].
Recently several interesting propositions on how to combine the binary classifiers were proposed. Wu et al. [121] used pairwise coupling, Friedman employed Max-Win rule [122],
9
Hullermeier proposed the adaptive weighted voting procedure [123]. A comprehensive recent
survey of binary classifier ensembles is [124]. It worth mentioning the one-class classification
model which is the special case of binary classifier trained in the absence of counterexamples. Its
main goal is to model normality in order to detect anomaly or outliers from the target class [125].
To combine such classifiers the typical methods developed for binary ones are used [126] but it is
worth mention the work by Wilk and Wozniak where authors restored multi-class classification
task using a pool of one-class classifiers and the fuzzy inference system [127]. The combination
methods dedicated the one-class classifiers still await a proper attention [128].
3.2.3. Diversifying models
Ensembles with individual classifiers based on different classification models take advantage
of the different biases of each classifier model [3]. However, the combination rule should be
carefully chosen. We can combine the class labels but in the case of continuous outputs we
have to normalize them e.g., using fuzzy approach [127]. We could use the different versions
of the same model as well, because many machine learning algorithms do not guarantee to find
the optimal classifier. Combining the results of various initializations may give good results.
Alternatively, a pool of classifiers can be produced by noise injection. Regarding neural networks
[129] it is easy to train pools of networks where each of them is trained starting from randomly
chosen initial weights. Regarding decisions tree we can choose randomly the test for a given
node among the possible tests according to the value of a splitting criterion.
4. Fuser design
Some works consider the answers from a given Oracle as the reference combination model
[130]. The Oracle is an abstract combination model, built such that if at least one of the individual
classifiers provides the correct answer, then the MCS committee outputs the correct class too.
Some researches used the Oracle in comparative experiments to provide a performance upper
bound for classifier committee [10] or information fusion methods [131]. A simple example
shows the risks of the Oracle model: assume we have two classifiers for a binary class problem,
a random one and the other that always returns the opposite decision; hence the Oracle will
always return the correct answer. As a consequence the Oracle model does not fit in the Bayesian
paradigm. Raudys [132] noticed that Oracle is a kind of quality measure of a given individual
classifier pool. Let us systematize methods of classifier fusion, which on the one hand could use
class labels or support function, on the other hand combination rules could be given or be the
results of training. The taxonomy of decision fusion strategies is depicted in Fig. 5.
4.1. Class label fusion
Early algorithms performing fusion of classifier responses [9; 10; 61] only implemented
majority voting schemes in three main versions [22]:
unanimous voting, so that the answer requires that all classifiers agree,
simple majority, so that the answer is given if majority is greater than half the pool of
classifiers,
majority voting, taking the answer with the highest number of votes.
10
NONTRAINABLE
TRAINABLE
(separately or co-trained)
BASED ON CLASSLABELS
Figure 5: A taxonomy of fusing strategies for the combination of MCS individual decisions.
The expected error of majority voting (for independent classifiers with the same quality) was
estimated in 1794 according to Bernoullis equation, proven as the Condorcet Jury Theorem
[5]. Later works focused on the analytically derived classification performance of combined
classifiers hold only when strong conditions are met [8] so that they are not useful from practical
point of view. Alternative voting methods weight differently the decisions coming from different
committee members [22; 133]. The typical architecture of combined classifier based on class
labels is presented in the left diagram of Fig. 6. In [134] authors distinguished the types of
weighted voting according to the classifier, both to the classifier and the class, and, finally, to
features values, the classifier and the class. Anyway, no one of these models can improve over
the Oracle. To achieve that we need additional information, such as the feature values [132; 135;
136] as depicted in the right diagram of Fig. 6.
classifer 1
classifer 1
classifier 2
object
decisi
on
decision
fuser
(combination rule)
.
.
.
classifier n
classifier 2
object
decision
decisi
on
decision
.
.
.
classifier n
classifier
decision
decision
decision
features
Figure 6: Architecture of the MCS making decision on the basis of class label fusion only (left diagram). The right
diagram corresponds to a MCS using additional information from the feature values.
the optimal projective fuser of [140], the combination of neural networks outputs according to
their accuracy [141], and Nave Bayes as the MCS combination method [142].
object
classifer 1
supports for
each class
classifier 2
supports for
each class
.
.
.
classifier n
count common
supports for each
class and make
decision according
to them
decision
supports for
each class
Figure 7: Architecture of the MCS which computes the decision on the basis of support function combination.
Some analytical properties and experimental evaluations of aggregating methods were presented in [10; 31; 143; 144]. The aggregating methods use simple operators such as supremum
or the mean value. They do not involve learning. However, they have little practical applicability
because of the hard conditions imposed by them [145]. The main aggregating advantage is that
it counteracts over-fitting of individual classifiers. According to [134], the following types of
weighted aggregation can be identified depending on: (a) only the classifier id, (b) the classifier
and the feature vector, (c) on the classifier and the class, and (d) on the classifier, the class, and
the feature vector. For two-class recognition problems only the last two types of aggregation allow to produce compound classifier which may improve the Oracle. For many-class problems, it
is possible to improve the Oracle [131] using any of these aggregation methods. Finally, another
salient approach is the mixture of experts [146; 147] which combines classifier outputs using
so-called input dependent gating function. Tresp and Taniguchi [148] proposed a linear function
for this fuser model, and Cheeseman [149] proposed a mixture of Gaussian.
4.3. Trainable Fuser
Fuser weight selection can be treated as a specific learning process [31; 136]. Shlien [150]
used Dempster and Shafers theory to reach a consensus on the weights to combine decision
trees. Wozniak [151] trained the fuser using perceptron-like learning, evolutionary algorithm
[152; 153]. Zheng used data envelopment analysis [154]. Other fuser trainable methods may be
strictly related to ensemble pruning methods, when authors use some heuristic search algorithm
to select the classifier ensemble, as [72; 141] according to the chosen fuser.
We have to mention the group of combination methods built from pools of heterogenous
classifiers, i.e. using different classification models, such as stacking [155]. This method trains
combination block using individual classifier outputs presented during classification of the whole
training set. Most of the combination methods do not take into consideration possible relations among individual classifiers. Huang and Suen [156] proposed Behavior-Knowledge Space
method which aggregates the individual classifiers decision on the basis of the statistical approach.
5. Concept drift
Before entering the discussion of practical applications we consider a very specific topic
of real life relevance which is known as Concept Drift in knowledge engineering domains, or
non-stationary processes in signal processing and statistics domains. Most of the conventional
12
classifiers do not take into consideration this phenomenon. Concept Drift means that the statistical dependencies between object features and its classification may change in time, so that future
data may be badly processed if we maintain the same classification, because the object category
or its properties will be changing. Concept drift occurs frequently in real life [157]. MCS are
specially well suited to deal with Concept Drift.
Machine learning methods in security applications (like spam filters or IDS/IPS) [158] or
decision support systems for marketing departments [159] require to take into account new training data with potentially different statistical properties[116]. The occurrence of Concept Drift
decreases the true classification accuracy dramatically. The most popular approaches are the
Streaming Ensemble Algorithm (SEA) [111] and the Accuracy Weighted Ensemble (AWE)[160].
Incoming data are collected in data chunks, which are used to train new models. The individual
classifiers evaluation is done on their accuracy on the new data. The best performing classifiers are selected to constitute the MCS committee in the next time epoch. As the decision rule,
the SEA uses a majority voting, whereas the AWE uses a weighted voting strategy. Kotler et
al. present the Dynamic Weighted Majority (DWM) algorithm [114] which modifies the decision combination weights and updates the ensemble according to number of incorrect decisions
made by individual classifiers. When a classifier weight is too small, then it is removed from the
ensemble, a new classifier is trained and added to the ensemble in its place.
A difficult problem is drift detection, which is the problem of deciding that the Concept
Drift has taken place. The current research direction is to propose an additional binary classifier
giving the decision to rebuild the classifiers. The drift detector can be based on changes in the
probability distribution of the instances [161163] or classification accuracy [164; 165]. Not
all classification algorithms dealing with concept drift require drift detection, because they can
adjust the model to incoming data [166][? ].
6. Applications
Reported applications of classifier ensembles have grown astoundingly in the recent years
due to the increase in computational power allowing training of large collections of classifiers
in practical application time constraints. A recent review appears in [18]. Sometimes the works
combine diverse kinds of classifiers, so-called heterogeneous MCS. Homogeneous MCS, such as
Random Forest (RF), are composed of classifiers of the same kind. In the works revised below,
basic classifiers are Multi-Layer Perceptron (MLP), k-Nearest Neighbor (kNN), Radial Basis
Function (RBF), Support Vector Machines (SVM), Probabilistic Neural Networks (PNN), and
Maximum Likelihood (ML) classifiers.
We review in this section recent applications to remote sensing data, computer security, financial risk assessment, fraud detection, recommender systems, and medical computer aided
diagnosis.
6.1. Remote sensing
The main problems addressed by MCS in remote sensing domains are the land cover mapping
and change detection. Land cover mapping consists in the identification of materials that are in
the surface of the area being covered. Depending on the application, a few general classes may be
identified, i.e. vegetation, water, buildings, roads, or a more precise classification can be required,
i.e. identifying tree or crop types. Applications include agriculture, forestry, geology, urban
planning, infrastructure degradation assessment. Change detection consists in the identification
13
of places where the land cover has changed in time, it implies the computation over time series
of images. Change detection may or may not be based on previous or separate land cover maps.
Remote sensing classification can be done on a variety of data sources, sometimes performing
fusion of different data modalities. Optical data has better interpretability by humans, but land
is easily occluded by weather conditions, i.e. cloud formations. Hyperspectral sensing provides
high-dimensional data at each image pixel, with high spectral resolution. Synthetic Aperture
Radar (SAR) is not affected by weather or other atmospheric conditions, so that observations are
better suited for continuous monitoring of seasonally changing land covers. SAR can provide
also multivariate data from varying radar frequencies. Other data sources are elevation maps,
and other ancillary information, such as the measurements of environmental sensors.
Land cover mapping
Early application of MCS to land cover mapping consisted in overproducing a large set of
classifiers and searching for the optimal subset [38; 65; 167]. To avoid the combinatorial complexity, the approach performs clustering of classifier error, aggregating similar classifiers. The
approach was proven to be optimal under some conditions on the classifiers. Interestingly, testing was performed on multi-source data, composing the pixels feature vector of joining multispectral with radar data channels, to compute the land cover map. The MCS was heterogenous,
composed of MLP, RBF, and PNN.
The application of RF to processing remote sensing data has been abundant in the literature. It has been applied to estimate land cover on Landsat data over Granada, Spain [168] and
multi-source data in a Colorado mountainous area [169]. Specifically, Landsat Multi-Spectral,
elevation, slope and aspect data are used as input features. The RF approach is able to successfully fuse these inhomogeneous informations. Works on hyperspectral images acquired by
the HyMap sensor have been addressed to build vegetation thematic maps [170], comparing RF
and decision tree-based Adaboost, as well as two feature selection methods: the out-of-bag and
a best-first search wrapper feature subset selection method. Diverse feature subsets are tested,
and the general conclusion is that tree ecotopes are better discriminated than grass ecotopes.
Further work with RF has been done assessing the uncertainty in modeling the distribution of
vegetation types [171], performing classification on the basis of environmental variables, in an
approach that combines spatial distribution modeling by spatial interpolation, using sequential
Gaussian simulation and the clustering of species into vegetation types. Dealing with labeled
data scarcity, there are methods [172] based on the combination of RF and the enrichment of the
training dataset with artificially generated samples in order to increase classifier diversity, which
is applied to Landsat multispectral data. Artificial data is generated from the Gaussian modeling of the data distribution. The application of RF to SAR multitemporal data aims to achieve
season invariant detection of several classes of land cover, i.e. grassland, ceral, forest, etc.[173].
RF performed best, with lowest spatial variability. Images were coregistered and some model
portability was tested, where the model trained on one SAR image was applied on other SAR
images of the same site obtained at different times. The success of RF for remote sensing images
has prompted the proposal of an specific computational environment [174].
Ensambles of SVM have been also applied to land cover map. Indeed, the ground truth
data scarcity has been attacked by an active learning approach to semi-supervised SVM training
[175]. The active learning approach is based on the clustering of the unlabeled data samples according to the clustering of the SVM outputs on the current training dataset. Samples with higher
membership coefficient are added to the corresponding class data, and the classifier is retrained
in an iterative process. These semi-supervised SVM are combined in a majority voting ensemble
14
and applied to the classification SPOT and Landsat optical data. Land cover classification in the
specific context of shallow waters has the additional difficulties of the scattering, refraction and
reflection effects introduced by the water cover. A robust process combines a parallel and a serial
architecture [176], where initial classification results obtained by SVM are refined in a second
SVM classifier and the final result is given by a linear combination of two ensembles of SVM
classifiers and a minimum distance classifier. Besides, the system estimates the water depth by
a bathymetry estimation process. The approach is applied to Landsat images for the estimation
of coral population in coastal waters. Polarimetric SAR data used for the classification of Boreal
forests require an ensemble of SVM [177]. Each of the SVM is specifically tuned to a class, with
specific feature selection process. Best results are obtained when multi-temporal data is used,
joining two images from two different seasons (summer and winter) and performing the feature
selection and training on the joint data vectors.
Change detection
Early application of MCS to land cover change detection was based on non-parametric algorithms, specifically MLP, k-NN, RBF, and ML classifiers [178; 179] , where classifier fusion
was performed either by majority voting, Bayesian average and maximum a posteriori probability. Testing data were Thematic Mapper multispectral images, and the Synthetic Aperture Radar
(SAR) of Landsat 5 satellite. Recent works on change detection in panchromatic images with
MCS follow three different decision fuser strategies: majority voting, Dempster-Shafer evidence
theory, and the Fuzzy Integral [180] . The sequential process of the images previous to classification includes pan-sharpening of the multi-temporal images, co-registration, raw radiometric
change detection by image subtraction and automatic thresholding, and a final MCS decision
computed on the multi-spectral data and the change detection data obtained from the various
pan-sharpening approaches.
6.2. Computer Security
Computer security is at the core of most critical services nowadays, from universities, banking, companies, communication. Secure information processing is a growing concern, and the
machine learning approaches are trying to provide predictive solutions that may allow to avoid
the negative impact of such attacks. Here we introduce some of the problems, with current solutions proposed from the MCS paradigm.
Distributed denial of service
Distributed denial of service (DDoS) are among the most threatening attacks that an Internet Service Provider may face. Distributed service providers, such as military applications,
e-healthcare and e-governance can be very sensitive to this type of attacks, which can produce
network performance degradation, service unavailability, and revenue loss. There is a need for
intelligent systems able to discriminate legitimate flash crowds from an attack. A general architecture for automatic detection of DDoS attacks is needed where the attack detection may be
performed by a MCS. The MCS constituent classifiers may be ANNs trained with robust learning algorithms, i.e. Resilient BackPropagation (RBP). Specifically, a boosting strategy is defined
on the ensemble of RBP trained ANNs, and a Neyman Pearson approach is used to make the
final decision [181]. This architecture may be based on Sugeno Adaptive Neuro-Fuzzy Inference
Systems (ANFIS) [182] . A critical issue of the approach is the need to report validation results,
which can only be based on recorded real life DDoS attacks. There are some public available
15
datasets to perform and report these results. However, results reported on these datasets may
not be informative of the system performance on new attacks which may have quite different
features. This is a pervasive concern in all security applications of machine learning algorithms.
Malware
Malicious code, such as trojans, virus, spyware, detection by anti-virus approaches can only
be performed after some instance of the code has been analyzed finding some kind of signature, therefore some degree of damage has already been done. Predictive approaches based on
Machine Learning techniques may allow anticipative detection at the cost of some false positives. Classifiers learn patterns in the known malicious codes extrapolating to yet unseen codes.
A taxonomy of such approaches is given in [183]. describing the basic code representation by
byte and opcode n-grams, strings, and others like portable executable features. Feature selection
processes, such as the Fisher score, are applied to find the most informative features. Finally,
classifiers tested in this problem include a wide variety of MCS combining diverse base classifiers with all standard fuser designs. Results have been reported that MCS overcome other
approaches, are better suitable for active learning needed to keep the classifiers updated and
tuned to the changing malicious code versions.
Intrusion detection
Intrusion Detection and Intrusion Prevention deal with the identification of intruder code in
a networked environment via the monitoring of communication patterns. Intruder detection performed as an anomaly detection process allows to detect previously unseen patterns, at the cost
of false alarms, contrary to signature based approaches. The problem is attacked by modular
MCS whose compounding base classifiers are one-class classifiers built by the Parzen window
probability density estimation approach [128]. Each module is specialized in a specific protocol or network service, so that different thresholds can be tuned for each module allowing some
optimization of the false alarm rate. On the other hand, Intrusion Prevention tries to impede the
execution of the intruder code by fail-safe semantics, automatic response and adaptive enforcement. An approach relies on the fact that Instruction Set Randomization prevents code injection
attacks, so that detected injected code can be used for adaptation of the anomaly classifier and
the signature-based filtering [184]. Clustering of n-grams is performed to obtain a model of the
normal communication behavior which is accurate allowing zero-day detection of worm infection even in the case of low payload or slow penetration [185]. The interesting proposed hybrid
intrusion detection was presented in [186], where decision trees and support vector machines are
combined as a hierarchical hybrid intelligent system model.
Wireless sensor networks
Wireless sensor networks (WSN) are collections of inexpensive, low power devices deployed
over a geographical space for monitoring, measuring and event detection. Anomalies in the WSN
can be due to failures in software or hardware, or to malicious attacks compelling the sensors to
bias or drop their information and measurements. Anomaly detection in WSN is performed using
an ensemble of binary classifiers, each tuned on diverse parameters and built following a different
approach (Average, autorregresive, neural network, ANFIS). The decision is made by a weighted
combination of the classifiers outputs [187].
16
this problem a combination of bagging and random subspace feature selection using SVM as
the base classifier has been developed and tested. The resulting method has increased diversity
improving results over a dataset provided by the Bank of China [195]. Bankruptcy prediction
is a dramatic special case of credit risk. Ensemble systems with diversity ensured by genetic
algorithm based selection of component classifiers is proposed in [196] for bankruptcy prediction
in South Korean firms. The prediction of failure of dotcom companies has been a matter of
research since the bubble explosion after the year two thousand. Tuning a hybrid of PNN, MLP
and genetic programming classifiers over a set of features selected applying a t-test and F-test
for relevance to the categorical variable has given some solutions [197]. The same approach is
reported in [198] to detect fraud in the financial statement of big companies.
Financial risks
Uncertainty in the financial operations is identified with the financial risks such as credit,
business, investment, and operational risks. Financial distress can be detected by clustering and
MCS in four different combination models. Clustering is performed by classical SOM and kmeans algorithms and used to partition the data space prior to MCS training [199] . Experimental
framework for the evaluation of financial risk assessment models, giving a specific performance
measures allow the exploration of computational solutions to these problems [200]. Several conventional classifiers and MCS have been tested in this framework using a large pool of datasets.
Bank performance and bankruptcy prediction is addressed using a widely heterogenous MCS
including PNN,RBF, MLP, SVM, CART trees, and a fuzzy rule system. The effect of PCA initial
dimensionality reduction is also tested [201]. The effect of feature construction from previous
experience and a priori information in the efficiency of classifiers for early warning of bank
failures is reported in [202].
New fraud trends
Prescription fraud has been identified as a cause of substantial monetary loss in health care
systems, it consists in the prescription of unnecessary medicaments. The research works need to
real life data from a large multi-center medical prescription database [203]. The authors use a
novel distance based on data-mining approach in a system which is capable of self-learning by
regular updates. The system is designed to perform on-line risky prescription detection followed
by off-line expert evaluation.
A new brand of frauds appear in the online gaming and lotteries, i.e. intended for money
laundering, whose detection is dealt with a mixture of supervised and unsupervised classifiers
[204]. To be adaptive to fraudster evolving strategies, it is required to emphasize online learning, and online cluster detection. Fraud in telecommunication systems involving usage beyond
contract specifications is dealt with in [205] by a preprocessing, clustering and classification
pipeline. Clustering has been found to improve classification performance, and boosted trees are
the best performing approach. The analysis of social networks by means of MCS may allow
the detection of fraud in automobile insurance, consisting in staging traffic accidents and issuing
fake insurance claims to their general or vehicle insurance company [206].
6.4. Medicine
Medicine is a big area of application of any innovative computational approach, dealing
with massive amounts of data in some instances, and with very imprecise or ambiguous data
in other situations. The range of applications is quite big, so here we only give a scrap of all
18
the current problems and approaches related with the MCS paradigm. In Medicine, a specific
research area since the inception of Artificial Intelligence is the construction of Computer Aided
Diagnosis (CAD) systems or Clinical Decision Support Systems (CDSS) [207], which involve as
the final step some kind of classifier predicting the subjects disease or normal status. In CDSS
development, there are several steps such as the definition of the sensor providing the data, the
preprocessing of the data to normalize it and remove noise, the selection of features, and the final
selection of the classifier.
Coronary diseases
A recent instance of CDSS is the application to cardiovascular disease diagnosis of an heterogenous collection of classifiers, composed of SVM, bayesian networks and ANN [208] finding ten new biomarkers. In this AptaCDSS-E process starts with the use of an aptamer biochip
scanning protein expression levels which is the input to physician taking the decisions afterwards.
Feature selection is performed by an ANOVA analysis. Doctor decisions are stored for system
retraining. Classifier combination is done by majority voting or hierarchical fusion. Many CAD
systems related with coronary diseases are based on the information provided by the electrocardiogram (ECG), so that many of them rely on the features extracted from them. Coronary artery
disease is a broad term that encompasses any condition that affects the heart. It is a chronic
disease in which the coronary arteries gradually harden and narrow, there have approaches to
provide CAD for this condition, such as the use of a mixture of three ANNs for the prediction
of coronary artery disease [209]. The dysfunction or abnormality of one or more of the heart
four valves is called valvular heart disease. Its diagnosis is performed by neural network ensembles in [209; 210] over features selected by a correlation analysis with the categorical variable.
Two separate ANNs are trained to identify myocardial infarction on training sets with different
statistics regarding the percentage of patients in [211]. The network specialized in healthy controls is applied to the new data, if the output is below a threshold the subject is deemed healthy,
otherwise the disease-specific network is applied to decide.
Proteomics
Proteins are said to have a common fold if they have the same major secondary structure in the
same arrangement and with the same topology. Machine learning techniques have been proposed
for three-dimensional protein structure prediction. Early approaches consisted in hybrid systems,
such as the ANN, statistical classifier and case base reasoning classifier combined by majority
voting of [212]. For instance, an ensemble of K-local hyperplanes based on random subspace
and feature selection has been tested [213], where feature selection is done according to distance
to the class centroids. A recent approach is the MarFold [214] combining by majority voting
three margin-based classifiers for protein fold recognition: the adaptive local hyperplane (ALH),
the k-neighborhood ALH and the SVM.
Neuroscience
In the field of Neurosciences, the machine learning approach is gaining widespread acceptation. It is used for the classification of image data searching for predictive non-invasive biomarkers that may allow early or prodromal diagnosis of a number of degenerative diseases which
have increasing impact in the society due to the aging of populations around the world. Diverse
MCS approaches have been applied to structural MRI data, specifically for the classification of
Alzheimer disease patients, such as an RVM based two stage pipeline [45], variations of Adaboost [215], hybridizations of kernel and Dendritic Computing approaches [216]. Classifier
19
Ensembles have been applied to the classification of fMRI data [217; 218] and its visual decoding [219], which is the reconstruction of the visual stimuli from the fMRI data.
6.5. Recommender systems
Nowadays, recommender systems are the focus of intense research [220]. They try to help
consumers to select the product that may be interesting for them based on their previous searches
and transactions, but such systems are expanding beyond typical sales. They are used to predict
which mobile telephone subscribers are in risk of switching to another provider, or to advice
conference organizers about assigning papers to peer reviewers [221]. Burke [222] proposed
hybrid recommender systems combining two or more recommendation techniques to improve
performance avoiding the drawbacks of an individual recommender. Similar observations were
confirmed by Balabanovic et al. [223] and Pazzani [224] who demonstrated that hybrid method
recommentations improve collaborative and content-based approaches.
There are several interesting works which apply the hybrid and combined approach to recommender systems. Jahrer and Toscher [225] demonstrated the advantage of ensemble learning applied to the combination of different collaborative filtering algorithms on the Netix Prize dataset.
Porcel et al. [226] developed an hybrid fuzzy recommender system to help disseminate information about research resources in the field of interest of a user. Claypool et al. [227] performed a
linear combination of the ratings obtained from individual recommender systems into one final
recommendation, while Pazzani proposed to use a voting scheme [224]. Billsus and Pazzani
[228] selected the best recommendation on the basis of a recommendation quality metric as the
level of confidence while Tran and Cohen [229] preferred an individual which is the most consistent with the previous ratings of the user. Kunaver et al. [230] proposed Combined Collaborative
Recommender based on three different collaborative recommender techniques. Goksedef and
Gundoz-Oguducu [231] combined the results of several recommender techniques based on Web
usage mining.
7. Final remarks
We have summarized the main research streams on multiple classifier systems, also known in
the literature as combined classifier or classifier ensemble. Such hybrid systems are the focus of
intense research recently, so fruitful that our review could not be exhaustive. Key issues related
to the problem under consideration are classifier diversity and methods of classifier combination.
The diversity is believed to provide improved accuracy and classifier performance. Most works
try to obtain maximum diversity by different means: introducing classifier heterogeneity, bootstrapping the training data, randomizing feature selection, randomizing subspace projections,
boosting the data weights, and many combinations of these ideas. Nowadays, the diversity hypothesis has not been fully proven, either theoretically or empirically. However, the fact is that
MCSs show in most instances improved performance, resilience and robustness to high data dimensionality and diverse forms of noise, such as labeling noise.
The there are several propositions how to combine the classifier outputs, what was presented in
this work, nonetheless we point out that classifier combination is not the only way to to produce
hybrid classifier systems. We envisage further possibilities of hybridization such as
Merging the raw data from different sources into one repository and then train the classifier.
Merging the raw data and a prior expert knowledge (e.g., learning sets and human expert
rules to improve rules on the basis of incoming data).
20
Merging a prior expert knowledge and classification models returned by machine learning
procedures.
For such a problem we have to take into consideration issues related to data privacy, computational and memory efficiency.
8. Acknowledgment
We would like to thank the anonymous reviewers for their diligent work and efficient efforts.
We are also grateful to the Editor-in-Chief, Prof. Belur V. Dasarathy, who encouraged us to write
this survey for this prestigious journal.
Micha Wozniak was supported by The Polish National Science Centre under the grant N N519
576638 which is being realized in years 2010-2013.
References
[1] J. Neumann, The computer and the brain, Yale University Press, New Haven, CT, USA, 1958.
[2] A. Newell, Intellectual issues in the history of artificial intelligence, in: F. Machlup, U. Mansfield (Eds.), The study
of information: interdisciplinary messages, John Wiley & Sons, Inc., New York, NY, USA, 1983, pp. 187294.
[3] D. Wolpert, The supervised learning no-free-lunch theorems, in: In Proc. 6th Online World Conference on Soft
Computing in Industrial Applications, 2001, pp. 2542.
[4] C. K. Chow, Statistical independence and threshold functions, Electronic Computers, IEEE Transactions on EC14 (1) (1965) 66 68.
[5] L. Shapley, B. N.G., Optimizing group judgmental accuracy in the presence of interdependencies, Public Choice
43 (3) (1984) 32933.
[6] B. Dasarathy, B. Sheela, A composite classifier system design: Concepts and methodology, Proceedings of the
IEEE 67 (5) (1979) 708 713.
[7] L. Rastrigin, R. H. Erenstein, Method of Collective Recognition, Energoizdat, Moscow, 1981.
[8] L. Hansen, P. Salamon, Neural network ensembles, Pattern Analysis and Machine Intelligence, IEEE Transactions
on 12 (10) (1990) 993 1001. doi:10.1109/34.58871.
[9] L. Xu, A. Krzyzak, C. Suen, Methods of combining multiple classifiers and their applications to handwriting
recognition, Systems, Man and Cybernetics, IEEE Transactions on 22 (3) (1992) 418 435.
[10] K. Tumer, J. Ghosh, Analysis of decision boundaries in linearly combined neural classifiers, Pattern Recognition
29 (2) (1996) 341 348.
[11] T. Ho, J. J. Hull, S. Srihari, Decision combination in multiple classifier systems, IEEE Transactions on Pattern
Analysis and Machine Intelligence 16 (1) (1994) 6675.
[12] L. Breiman, Bagging predictors, Machine Learning 24 (2) (1996) 123140.
[13] R. Schapire, The strength of weak learnability, Mach. Learn. 5 (2) (1990) 197227.
[14] Y. Freund, Boosting a weak learning algorithm by majority, Information Computing 121 (2) (1995) 256285.
[15] M. Kearns, U. Vazirani, An introduction to computational learning theory, MIT Press, Cambridge, MA, USA,
1994.
[16] D. Angluin, Queries and concept learning, Machine Learning 2 (4) (1988) 319342.
[17] A. Jain, R. Duin, M. Jianchang, Statistical pattern recognition: a review, Pattern Analysis and Machine Intelligence, IEEE Transactions on 22 (1) (2000) 4 37.
[18] N. Oza, K. Tumer, Classifier ensembles: Select real-world applications, Information Fusion 9 (1) (2008) 420.
[19] R. Polikar, Ensemble based systems in decision making, IEEE Circuits and Systems Magazine 6 (3) (2006) 2145.
[20] R. Polikar, Ensemble learning, Scholarpedia 3 (12) (2008) 2776.
[21] L. Rokach, Taxonomy for characterizing ensemble methods in classification tasks: A review and annotated bibliography, Computational Statistics and Data Analysis 53 (12) (2009) 4046 4072.
[22] L. Kuncheva, Combining Pattern Classifiers: Methods and Algorithms, Wiley-Interscience, 2004.
[23] L. Rokach, Pattern classification using ensemble methods, Series in machine perception and artificial intelligence,
World Scientific, 2010.
[24] G. Seni, J. Elder, Ensemble Methods in Data Mining: Improving Accuracy Through Combining Predictions,
Morgan and Claypool Publishers, 2010.
21
[25] B. Baruque, E. Corchado, Fusion Methods for Unsupervised Learning Ensembles, Springer-Verlag New York,
Inc., 2011.
[26] R. Duda, P. Hart, D. Stork, Pattern Classification, 2nd Edition, Wiley, New York, 2001.
[27] E. Alpaydin, Introduction to Machine Learning, Second Edition, The MIT Press, 2010.
[28] C. Bishop, Pattern Recognition and Machine Learning (Information Science and Statistics), Springer-Verlag New
York, Inc., Secaucus, NJ, USA, 2006.
[29] T. Dietterich, Ensemble methods in machine learning, in: Multiple Classifier Systems, Vol. 1857 of Lecture Notes
in Computer Science, Springer Berlin, Heidelberg, 2000, pp. 115.
[30] G. Marcialis, F. Roli, Fusion of face recognition algorithms for video-based surveillance systems, G.L. Foresti, C.
Regazzoni, P. Varshney Eds (2003) 235250.
[31] S. Hashem, Optimal linear combinations of neural networks, Neural Networks 10 (4) (1997) 599614.
[32] R. Clemen, Combining forecasts: A review and annotated bibliography, International Journal of Forecasting 5 (4)
(1989) 559583.
[33] J. Quinlan, C4.5: Programs for Machine Learning, Morgan Kaufmann Series in Machine Learning, Morgan
Kaufmann Publishers, 1993.
[34] T. Wilk, M. Wozniak, Complexity and multithreaded implementation analysis of one class-classifiers fuzzy combiner, in: E. Corchado, M. Kurzynski, M. Wozniak (Eds.), Hybrid Artificial Intelligent Systems, Vol. 6679 of
Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2011, pp. 237244.
[35] T. Kacprzak, K. Walkowiak, M. Wozniak, Optimization of overlay distributed computing systems for multiple
classifier system - heuristic approach, Logic Journal of IGPLdoi:10.1093/jigpal/jzr020.
[36] K. Walkowiak, Anycasting in connection-oriented computer networks: Models, algorithms and results, Int. J.
Applied Mathematics and Computer Sciences 20 (1) (2010) 207220.
[37] R. Agrawal, R. Srikant, Privacy-preserving data mining, SIGMOD Records 29 (2) (2000) 439450.
[38] G. Giacinto, F. Roli, G. Fumera, Design of effective multiple classifier systems by clustering of classifiers, in:
Pattern Recognition, 2000. Proceedings. 15th International Conference on, Vol. 2, 2000, pp. 160 163 vol.2.
[39] T. Ho, Complexity of classification problems and comparative advantages of combined classifiers, in: Proceedings
of the First International Workshop on Multiple Classifier Systems, MCS 00, Springer-Verlag, London, UK, UK,
2000, pp. 97106.
[40] F. Roli, G. Giacinto, Design of Multiple Classifier Systems, World Scientific Publishing, 2002.
[41] L. Lam, Classifier combinations: Implementations and theoretical issues, in: Proceedings of the First International
Workshop on Multiple Classifier Systems, MCS 00, Springer-Verlag, London, UK, UK, 2000, pp. 7786.
[42] A. F. R. Rahman, M. C. Fairhurst, Serial combination of multiple experts: A unified evaluation, Pattern Analysis
and Applications 2 (1999) 292311.
[43] G. Fumera, I. Pillai, F. Roli, A two-stage classifier with reject option for text categorisation, in: 5th Int. Workshop
on Statistical Techniques in Pattern Recognition (SPR 2004), Vol. 3138, Springer, Springer, Lisbon, Portugal,
2004, pp. 771779.
[44] P. Bartlett, M. Wegkamp, Classification with a reject option using a hinge loss, J. Machine Learning Research 9
(2008) 18231840.
[45] M. Termenon, M. Grana, A two stage sequential ensemble applied to the classification of alzheimers disease
based on mri features, Neural Processing Letters 35 (1) (2012) 112.
[46] P. Clark, T. Niblett, The cn2 induction algorithm, Machine Learning 3 (4) (1989) 261283.
[47] R. Rivest, Learning decision lists, Machine Learning 2 (3) (1987) 229246.
[48] Y. Freund, R. Schapire, A decision-theoretic generalization of on-line learning and an application to boosting,
Journal of Computer and System Sciences 55 (1) (1997) 119139. doi:10.1006/jcss.1997.1504.
[49] X. Wu, V. Kumar, J. R. Quinlan, J. Ghosh, Q. Yang, H. Motoda, G. J. McLachlan, A. Ng, B. Liu, P. S. Yu, Z.-H.
Zhou, M. Steinbach, D. J. Hand, D. Steinberg, Top 10 algorithms in data mining, Knowledge and Information
Systems 14 (1) (2008) 137. doi:10.1007/s10115-007-0114-2.
[50] R. Schapire, The strength of weak learnability, Machine Learning 5 (2) (1990) 197227.
doi:10.1023/A:1022648800760.
[51] J. Kivinen, M. K. Warmuth, Boosting as entropy projection, in: Proceedings of the twelfth annual conference on
Computational learning theory, 1999.
URL http://dl.acm.org/citation.cfm?id=307424
[52] D. Partridge, W. Krzanowski, Software diversity: practical statistics for its measurement and exploitation, Information and Software Technology 39 (10) (1997) 707 717.
[53] G. Brown, L. Kuncheva, good and bad diversity in majority vote ensembles, in: Proceedings MCS 2010,
2010, pp. 124133.
[54] M. Smetek, B. Trawinski, Selection of heterogeneous fuzzy model ensembles using self-adaptive genetic algorithms, New Generation Computing 29 (2011) 309327.
[55] A. J. C. Sharkey, N. Sharkey, Combining diverse neural nets, Knowl. Eng. Rev. 12 (3) (1997) 231247.
22
[56] A. Krogh, J. Vedelsby, Neural network ensembles, cross validation, and active learning, in: Advances in Neural
Information Processing Systems, Vol. 7, 1995, pp. 231238.
[57] G. Zenobi, P. Cunningham, Using diversity in preparing ensembles of classifiers based on different feature subsets
to minimize generalization error, Machine Learning: ECML 2001 (2001) 576587.
[58] G. Brown, J. Wyatt, R. Harris, X. Yao, Diversity creation methods: a survey and categorisation, Information
Fusion 6 (1) (2005) 520.
[59] N. Ueda, R. Nakano, Generalization error of ensemble estimators, in: Proceedings of IEEE International Conference on Neural Networks., Washington, USA, 1996, pp. 9095.
[60] G. Brown, J. Wyatt, P. Tino, Managing diversity in regression ensembles, J. Machine Learning Research 6 (2005)
16211650.
[61] L. Kuncheva, C. Whitaker, C. Shipp, R. Duin, Limits on the majority vote accuracy in classifier fusion, Pattern
Analysis and Applications 6 (2003) 2231.
[62] Y. Bi, The impact of diversity on the accuracy of evidential classifier ensembles, International Journal of Approximate Reasoning 53 (4) (2012) 584607.
[63] D. Margineantu, T. Dietterich, Pruning adaptive boosting, in: Proceedings of the Fourteenth International Conference on Machine Learning, ICML 97, Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1997, pp.
211218.
[64] D. Skalak, The sources of increased accuracy for two proposed boosting algorithms, in: In Proc. American Association for Arti Intelligence, AAAI-96, Integrating Multiple Learned Models Workshop, 1996, pp. 120125.
[65] G. Giacinto, F. Roli, Design of effective neural network ensembles for image classification purposes, Image Vision
Computing 19 (9-10) (2001) 699707.
[66] R. Kohavi, D. Wolpert, Bias plus variance decomposition for zero-one loss functions, in: ICML-96, 1996.
[67] J. Fleiss, J. Cuzick, The reliability of dichotomous judgments: unequal numbers of judgments per subject, Applied
Psychological Measurement 4 (3) (1979) 537542.
[68] P. Cunningham, J. Carney, Diversity versus quality in classification ensembles based on feature selection, in:
Proceedings of the 11th European Conference on Machine Learning, ECML 00, Springer-Verlag, London, UK,
UK, 2000, pp. 109116.
[69] C. Shipp, L. Kuncheva, Relationships between combination methods and measures of diversity in combining
classifiers, Information Fusion 3 (2) (2002) 135148.
[70] E. K. Tang, P. N. Suganthan, X. Yao, An analysis of diversity measures, Machine Learning 65 (1) (2006) 247271.
[71] G. Martinez-Mu/ noz, D. Hern/andez-Lobato, A. Suarez, An analysis of ensemble pruning techniques based on
ordered aggregation, IEEE Transactions on Pattern Analysis and Machine Intelligence 31 (2) (2009) 245259.
[72] D. Ruta, B. Gabrys, Classifier selection for majority voting, Information Fusion 6 (1) (2005) 63 81.
[73] R. Banfield, L. Hall, K. Bowyer, W. Kegelmeyer, Ensemble diversity measures and their application to thinning,
Information Fusion 6 (1) (2005) 4962.
[74] Z.-H. Zhou, J. Wu, W. Tang, Ensembling neural networks: Many could be better than all, Artificial Intelligence
137 (1-2) (2002) 239263.
[75] B. Gabrys, D. Ruta, Genetic algorithms in classifier fusion, Applied Soft Computing 6 (4) (2006) 337347.
[76] I. Partalas, G. Tsoumakas, I. Vlahavas, Pruning an ensemble of classifiers via reinforcement learning, Neurocomputing 72 (7-9) (2009) 19001909.
[77] Q. Dai, A competitive ensemble pruning approach based on cross-validation technique, Knowledge-Based Systems (0) (2012) . doi:10.1016/j.knosys.2012.08.024.
[78] Y. Peng, Q. Huang, P. Jiang, J. Jiang, Cost-sensitive ensemble of support vector machines for effective detection of microcalcification in breast cancer diagnosis, in: L. Wang, Y. Jin (Eds.), Fuzzy Systems and Knowledge
Discovery, Vol. 3614 of Lecture Notes in Computer Science, Springer Berlin / Heidelberg, 2005, pp. 483483.
[79] K. Jackowski, B. Krawczyk, M. Woniak, Cost-sensitive splitting and selection method for medical decision support system, in: H. Yin, J. A. Costa, G. Barreto (Eds.), Intelligent Data Engineering and Automated Learning IDEAL 2012, Vol. 7435 of Lecture Notes in Computer Science, Springer Berlin Heidelberg, 2012, pp. 850857.
[80] W. Du, Z. Zhan, Building decision tree classifier on private data, in: Proceedings of the IEEE international
conference on Privacy, security and data mining - Volume 14, CRPIT 14, Australian Computer Society, Inc.,
Darlinghurst, Australia, Australia, 2002, pp. 18.
[81] B. Krawczyk, M. Wozniak, Privacy preserving models of k-nn algorithm, in: R. Burduk, M. Kurzynski, M. Wozniak, A. Zolnierek (Eds.), Computer Recognition Systems 4, Vol. 95 of Advances in Intelligent and Soft Computing, Springer Berlin / Heidelberg, 2011, pp. 207217.
[82] Y. Lindell, B. Pinkas, Secure multiparty computation for privacy-preserving data mining, IACR Cryptology ePrint
Archive 2008 (2008) 197.
[83] K. Walkowiak, S. Sztajer, M. Wozniak, Decentralized distributed computing system for privacy-preserving combined classifiers modeling and optimization, in: B. Murgante, O. Gervasi, A. Iglesias, D. Taniar, B. Apduhan
(Eds.), Computational Science and Its Applications - ICCSA 2011, Vol. 6782 of Lecture Notes in Computer
23
24
25
26
[165] M. Baena-Garca, J. del Campo-Avila,
R. Fidalgo, A. Bifet, R. Gavalda, R. Morales-Bueno, Early drift detection
method, in: In Fourth International Workshop on Knowledge Discovery from Data Streams, 2006.
[166] I. Zliobaite, Change with delayed labeling: When is it detectable?, in: Proceedings of the 2010 IEEE International
Conference on Data Mining Workshops, ICDMW 10, IEEE Computer Society, Washington, DC, USA, 2010, pp.
843850.
[167] G. Giacinto, F. Roli, L. Bruzzone, Combination of neural and statistical algorithms for supervised classification
of remote-sensing images, Pattern Recognition Letters 21 (5) (2000) 385 397.
[168] V. Rodriguez-Galiano, B. Ghimire, J. Rogan, M. Chica-Olmo, J. Rigol-Sanchez, An assessment of the effectiveness of a random forest classifier for land-cover classification, ISPRS Journal of Photogrammetry and Remote
Sensing 67 (0) (2012) 93 104.
[169] P. Gislason, J. Benediktsson, J. Sveinsson, Random forests for land cover classification, Pattern Recognition
Letters 27 (4) (2006) 294 300.
[170] J.-W. Chan, D. Paelinckx, Evaluation of random forest and adaboost tree-based ensemble classification and spectral band selection for ecotope mapping using airborne hyperspectral imagery, Remote Sensing of Environment
112 (6) (2008) 2999 3011.
[171] J. Peters, N. Verhoest, R. Samson, M. Meirvenne, L. Cockx, B. Baets, Uncertainty propagation in vegetation
distribution models based on ensemble classifiers, Ecological Modelling 220 (6) (2009) 791 804.
[172] M. Han, X. Zhu, W. Yao, Remote sensing image classification based on neural network ensemble algorithm,
Neurocomputing 78 (1) (2012) 133 138.
[173] B. Waske, M. Braun, Classifier ensembles for land cover mapping using multitemporal sar imagery, ISPRS Journal
of Photogrammetry and Remote Sensing 64 (5) (2009) 450 457, theme Issue: Mapping with SAR: Techniques
and Applications.
[174] B. Waske, S. van der Linden, C. Oldenburg, B. Jakimow, A. Rabe, P. Hostert, imagerf- a user-oriented implementation for remote sensing image analysis with random forests, Environmental Modelling & Software 35 (0) (2012)
192 193.
[175] U. Maulik, D. Chakraborty, A self-trained ensemble with semisupervised svm: An application to pixel classification of remote sensing imagery, Pattern Recognition 44 (3) (2011) 615 623.
[176] A. Henriques, A. Doria-Neto, R. Amaral, Classification of multispectral images in coral environments using a
hybrid of classifier ensembles, Neurocomputing 73 (7-9) (2010) 1256 1264.
[177] Y. Maghsoudi, M. Collins, D. Leckie, Polarimetric classification of boreal forest using nonparametric feature
selection and multiple classifiers, International Journal of Applied Earth Observation and Geoinformation 19 (0)
(2012) 139 150.
[178] L. Bruzzone, R. Cossu, G. Vernazza, Combining parametric and non-parametric algorithms for a partially unsupervised classification of multitemporal remote-sensing images, Information Fusion 3 (4) (2002) 289 297.
[179] L. Bruzzone, R. Cossu, G. Vernazza, Detection of land-cover transitions by combining multidate classifiers, Pattern Recognition Letters 25 (13) (2004) 1491 1500.
[180] P. Du, S. Liu, J. Xia, Y. Zhao, Information fusion techniques for change detection from multi-temporal remote
sensing images, Information Fusion 14 (1) (2013) 19 27.
[181] P. Arun-Raj-Kumar, S. Selvakumar, Distributed denial of service attack detection using an ensemble of neural
classifier, Computer Communications 34 (11) (2011) 1328 1341.
[182] P. Kumar, S. Selvakumar, Detection of distributed denial of service attacks using an ensemble of adaptive and
hybrid neuro-fuzzy systems, Computer Communications (0) (2012) .
[183] A. Shabtai, R. Moskovitch, Y. Elovici, C. Glezer, Detection of malicious code by applying machine learning
classifiers on static features: A state-of-the-art survey, Information Security Technical Report 14 (1) (2009) 16
29.
[184] M. Locasto, K. Wang, A. Keromytis, S. Stolfo, Flips: hybrid adaptive intrusion prevention, in: Proceedings of
the 8th international conference on Recent Advances in Intrusion Detection, RAID05, Springer-Verlag, Berlin,
Heidelberg, 2006, pp. 82101.
[185] K. Wang, G. Cretu, S. Stolfo, Anomalous payload-based worm detection and signature generation, in: Proceedings
of the 8th international conference on Recent Advances in Intrusion Detection, RAID05, Springer-Verlag, Berlin,
Heidelberg, 2006, pp. 227246.
[186] S. Peddabachigari, A. Abraham, C. Grosan, J. Thomas, Modeling intrusion detection system using hybrid intelligent systems, J. Netw. Comput. Appl. 30 (1) (2007) 114132.
[187] D.-I. Curiac, C. Volosencu, Ensemble based sensing anomaly detection in wireless sensor networks, Expert Systems with Applications 39 (10) (2012) 9087 9096.
[188] R. J. Bolton, D. J. Hand, Statistical fraud detection: A review, Statistical Science 17 (3) (2002) 235255.
[189] F. Louzada, A. Ara, Bagging k-dependence probabilistic networks: An alternative powerful fraud detection tool,
Expert Systems with Applications 39 (14) (2012) 11583 11592.
[190] S. Bhattacharyya, S. Jha, K. Tharakunnel, J. Westland, Data mining for credit card fraud: A comparative study,
27
28
29
40000
combined classifier
35000
30000
classifier ensemble
25000
classifier fusion
hybrid classifier
20000
15000
10000
5000
0
till 1990
4000
combined classifier
3500
3000
classifier ensemble
2500
classifier fusion
hybrid classifier
2000
1500
1000
500
0
till 1990
2500
combined classifier
multiple classifier system
2000
classifier ensemble
classifier fusion
1500
hybrid classifier
1000
500
0
till 1990